ming topics.
This is directly opposed to the often made
assumption that a student's academic discipline is a good
predictor of potential competency in programming. This
result implies that if a course is suitably constructed and
presented, there is no need to segregate students from
different academic disciplines due to concerns based on
learning ability or interdisciplinary competitiveness.
5.2 Gender Differences
No significant difference was found in academic
performance between the genders. The only observable
difference is that women appear to be more uniform in
their academic performance than men.
5.3 Semester in School
The correlations found between semester in school
and academic performance were very low. This clearly
indicates that in a suitably constructed course, there is
no reason to worry more about those beginning their
university experience as opposed to those well along in
their academic careers. They all seem to succeed or fail
in much the same manner.
5.4 PAT Use
The results indicate that future programming skill is
not predictable by the most commonly used written test
(IBM's Programming Aptitude Test). The suggestion is
that this type of test should not be administered to college
level people as the results are unreliable.
6. Conclusions
It is not necessary to construct separate competing
courses for those from differing disciplines and levels of
academic experience as there is no apparent need to be
concerned with unequal capability. It is also not possible
to reliably predict success in learning programming on
the basis of normally observable external personal attri-
butes or standardized written tests when the people
involved are of college level.
Received May 1978; revised September 1978; accepted September 1979
References
!. Bateman, C.R. Predicting performance in a basic computer
course. Proc. of the Fifth Annual Meeting of the Amer. Inst. for
Decision Sciences, Boston, Mass., 1973.
2. Newstead, P.R. Grade and ability predictions in an introductory
programming course. SIGCSE Bull. 7, 2 (June 1975), pp. 87-91.
17
Programming and M.D. Mcllroy
Data Structures Editor
Minimal Perfect Hash
Functions Made
Simple
Richard J. Cichelli
Software Consulting Services,
Allentown, Pennsylvania
A method is presented for Computing machine
independent, minimal perfect hash functions of the
form: hash value <-- key length + the associated value
of the key's first character + the associated value of
the key's last character. Such functions allow single
probe retrieval from minimally sized tables of identifier
lists. Application areas include table lookup for
reserved words in compilers and filtering high frequency
words in natural language processing. Functions for
Pascal's reserved words, Pascal's predefined identifiers,
frequently occurring English words, and month
abbreviations are presented as examples.
Key Words and Phrases: hashing, hashing methods,
hash coding, direct addressing, dictionary lookup,
information retrieval, lexical analysis, identifier-to-
address transformations, perfect hashing functions,
perfect hash coding, scatter storage, searching, Pascal,
Pascal reserved words, backtracking
CR Categories: 3.7, 3.74, 4.34, 5.25, 5.39
Introduction
In several recent articles [2, 3] it has been asserted
that in general computing minimal perfect hash functions
for static identifier lists (keys) is difficult. In [1], Knuth
also notes the difficulty in computing perfect hash func-
tions. He estimates that for the set of words used here in
Example No. 3 a search for such a function might include
the examination of 1043 possibilities. A program imple-
Permission to copy without fee all or part of this material is
granted provided that the copies are not made or distributed for direct
commercial advantage, the ACM copyright notice and the title of the
publication and its date appear, and notice is given that copying is by
permission of the Association for Computing Machinery. To copy
otherwise, or to republish, requires a fee and/or specific permission.
Author's present address: R.J. Cichelli, Software Consulting Ser-
vices, 901 Whittier Drive, Allentown, PA 18103.
© 1980 ACM 0001-0782/80/0100-0017 $00.75.
Communications January 1980
of Volume 23
the ACM Number 1
menting the procedure outlined in this paper finds a
machine independent minimal perfect hash function for
these words in less than one second on a DEC PDP-11/
45 minicomputer. Other examples of such functions are
shown and an effective method for computing them is
described.
Although my algorithm requires exponential search
of potentially large spaces, it is based on intelligent
search of these spaces. The procedure has been shown to
find solutions for many sets in a few seconds. For thirty-
nine of Pascal's forty predefined identifiers (Example
No. 2), the potential search space was 2039 nodes. Com-
putation of the function required seven minutes on the
PDP-11/45. For lookup tables that will be used many
times, the method has proved quite practical. So far I
have encountered no instance of an interminably long
search.
The form of my hash function is:
Hash value ~ key length +
associated value of the key's first character +
associated value of the key's last character.
Example No. 1: Pascal's Reserved Words
For Pascal's 36 reserved words, the following list
defines the associated value for each letter:
A = II, B = 15, C = 1, D = 0 , E = 0 , F = 15, G = 3 , H = 15,1= 13,
J = 0 , K = 0 , L = 15, M = 15, N = 13, O = 0 , P = 15, Q = 0 , R = 14,
S = 6 , T = 6 , U = 14, V = 10, W = 6 , X = 0 , Y = 13, Z = 0.
(For lookup routines these values are stored in an integer
array indexed by the letters. Note: Associated values
need not be unique.)
The corresponding hash table with hash values run-
ning from 2 through 37 is as follows:
DO, END, ELSE, CASE, DOWNTO, GOTO, TO, OTHERWISE,
TYPE, WHILE, CONST, DIV, AND, SET, OR, OF, MOD, FILE,
RECORD, PACKED, NOT, THEN, PROCEDURE, WITH, RE-
PEAT, VAR, IN, ARRAY, IF, NIL, FOR, BEGIN, UNTIL, LABEL,
FUNCTION, PROGRAM.
As an example, consider the computation for
"CASE":
(1 ,--- "C") + (0 ~-- "E") + (4 ~ length("CASE")) = 5.
The advantage of hash functions of the above form
is that they are simple, efficient, and machine (i.e.,
character representation) independent. It is also likely
that any lexical scanning process will have, as a by-
product of its identifier scanning logic, the identifier
length and the values of the first and last characters.
Two disadvantages of functions of this form are: (1) it
requires that no two keys share length and first and last
characters and (2) for lists with more than about 45
items, segmentation into sublists may be necessary. (This
is a result of the limited range of hash values which the
functions produce.)
The associated values for each of the letters are
computed by the following procedure: (1) order the
identifier list, and (2) search, by backtracking, for a
solution.
18
The ordering process is twofold. First, order the keys
by the sum of the frequencies of the occurrences of each
key's first and last letter in the list. For example: "E"
occurs 9 times as a first or last letter in the Pascal reserved
word list. It is the most frequent letter and thus, "ELSE"
is the first word in the search list. "D" is the next most
frequent letter, and thus " E N D " is second. After the
words have been put in order by character occurrence
frequencies, modify the order of the list such that any
word whose hash value is determined by assigning the
associated character values already determined by pre-
vious words is placed next. Thus, after " O T H E R W I S E ''1
has been placed as the third element of the frequency
ordered list, the hash value of the word " D O " is deter-
mined and so it is placed fourth. (For example, during
search, after the placement of the word "END," a value
will be associated with "D," and after the placement of
the word "OTHERWISE," a value will be associated
with "O.") The ordering process causes inevitable hash
value conflicts to occur during search as early as possible,
thus pruning the search tree and speeding the computa-
tion.
The completely ordered search list for Pascal's re-
served words is:
ELSE, END, OTHERWISE, DO, DOWNTO, TYPE, TO, FILE, OF,
THEN, NOT, FUNCTION, RECORD, REPEAT, OR, FOR, PRO-
CEDURE, PACKED, WHILE, CASE, CONST, DIV, VAR, AND,
MOD, PROGRAM, NIL, LABEL, SET, IN, IF, GOTO, BEGIN,
UNTIL, ARRAY, WITH.
The backtracking search procedure then attempts to
find a set of associated values which will permit the
unique referencing of all members of the key word list.
It does this by trying the words one at a time in order.
The backtracking procedure is as follows: If both the
first and last letter of the identifier already have associ-
ated values, try the word. If either the first or last letter
has an associated value, vary the associated value of the
unassigned character from zero to the maximum allowed
associated value, trying each occurrence. If both letters
are as yet unassociated, vary the first and then the
second, trying each possible combination. (An exception
test is required to catch situations in which the first and
last letters are the same.) Each "try" tests whether the
given hash value is already assigned and, if not, reserves
the value and assigns the letters. If all identifiers have
been selected, print the solution and halt. Otherwise,
invoke the search procedure recursively to place the next
word. If the "try" fails, remove the word by backtrack-
ing.
The search time for computing such functions is
related to the number of identifiers to be placed, the
maximum value which is allowed to be associated with
a character, and the density of the resultant hash table.
1 Inclusion of the word "OTHERWISE" in Pascal's reserved word
list anticipates the acceptance by the Pascal Users Group of the
recommendation for a revised CASE construct submitted by its Inter-
national Working Group for Extensions.
Communications January 1980
of Volume 23
the ACM Number 1
If the table density is one (i.e., a minimal perfect hash)
and the maximum associated value is allowed to be the
count of distinct first and last letter occurrences (21 for
Pascal's reserved words), then the above procedure finds
a solution for Pascal's reserved words in about seven
seconds on a DEC PDP-11/45 using a straightforward
implementation of the algorithm in Pascal. (Without the
second ordering, the search required 5.5 hours.) If the
maximum associated value is limited to 15, as in the
above list, the search requires about 40 minutes. (There
is no solution with 14 as a maximum value.)
Incorporation of the above hash function into a
Pascal cross-reference program yielded a 10 percent
reduction in total run time for processing large programs.
The method replaced a well-coded binary search which
was used to exclude reserved words from cross-referenc-
ing.
Moral
This article does not have a conclusion, but it does
have a moral. In the words of the renowned chess
programmer, J. Gillogly, author of the Technology Chess
Program which was the prototype of the current gener-
ation of highly successful chess programs, "When all else
fails, try brute force."
Received May 1979; revised and accepted October 1979
References
1. Knuth, D.E. The Art of Computing Programming. Volume 3:
Sorting and Searching. Addison-Wesley, Reading, Mass., 1973, pp.
506-507.
2. Sheil, B.A. Median split trees: A fast lookup technique for
frequently occurring keys. Comm. A C M 21, 11 (Nov. 1978), 947-958.
3. Sprugnoli, R. Perfect hashing functions: A single probe retrieving
method for static sets. Comm. A C M 20, I l (Nov. 1977), 841-850.

Example No. 2
The second example is for the list of Pascal's prede-
fined identifiers.
A = 15, B = 9 , C = II, D = 19, E = 5 , F = 3 , G = O , H = O , I = 3 ,
J = 0 , K = 16, L = 13, M = I . N = 19, O = 0 , P = 18, Q = 0 , R = 0 ,
S = 15, T = 0 , U = 17, V = 0 , W = 10, X = 0, Y = 0, Z = 0.
GET, TEXT, RESET, OUTPUT, MAX1NT, INPUT, TRUE, INTE-
GER, EOF, REWRITE, FALSE, CHR, CHAR, TRUNC, REAL,
SQR, SQRT, WRITE, PUT, ORD, READ, ROUND, READLN,
EXP, PAGE, EOLN, COS, SUCC, DISPOSE, NEW, ABS, LN,
BOOLEAN, WRITELN, SIN, PACK, UNPACK, ARCTAN, PRED.

Computation of this function required about seven min-

utes. Note: Since the predefined identifier " O D D " con-
flicts with " O R D , " it was not included in the list.

Example No. 3: Frequently Occurring English Words

This example uses the word list o f [ 1, 2]. Search time
was less than one second.
A = 3 , B = 15, C = O , D = 7 , E = O , F = 15, G = 0 , H = 10,1 = 0,
J=O,K =O,L=O,M= t2, N = 1 3 , 0 = 7, P = O , Q = O , R = 12,
S = 6 , T = O , U = 15, V = 0 , W = 14, X = 0, Y = 9, Z = 0.
I, it, the, that, at, are, a, is, to, this, as, he, and, have, in, not, be, but,
his, had, or, on, was, of, her, by, you, with, which, for, from.

Example No. 4: Month Abbreviations

This example is from [3]. The function's form was
modified slightly to:
Hash value *-- associated value of the key's second character +
associated value of the key's third character.
A = 4 , B = 5, C = 2, D = O , E = O , F = O , G = 3, H = O , I = O ,
J=O,K =O,L--6, M=O,N=O, O = 5, P = I , Q = 0 , R = 6 ,
S=O,T=6, U=O,V=6, W=O,X=O,Y=5, Z=O.
JUN, SEP, DEC, AUG, JAN, FEB, JUL, APR, OCT, MAY, MAR,
NOV.

This form avoids the conflict between " J A N " and

" J U N " and takes into account the constant key length.
Search time was again well less than one second. Note:
The method presented here is applicable to sets up to
four times as large as those said to be feasible by the
methods described in [3].
19 Communications January 1980
of Volume 23
the ACM Number 1